How the Agent Thinks

A single trace transformed into a visual map of agent cognition. Each phase reveals a decision the agent made — what entered, what it remembered, what it concluded, and why it chose to stop.

1
Entry
Article enters the agent
IBM Think: "AI agents in 2025: expectations vs. reality" — a critical analysis of the gap between agent hype and enterprise readiness.
Source: ibm.com/think/insights/ai-agents-2025-expectations-vs-reality
2
Classify
The agent reads the article and classifies it
Before any analysis, the agent decides *what kind of thing* it's looking at. This shapes everything that follows.
domain: technology content_type: analysis author_stance: critic
Classification run before analysis. domain determines what past briefs to recall. content_type + author_stance determine the analysis lens.
3
Recall
Memory surfaces 3 past briefs
The agent searches its memory for relevant past thinking. These briefs are not just retrieved — they become active context that shapes the analysis.
2026-05-12
Organizations that start with customer needs and work backward to technology — rather than deploying AI capabilities and finding applications — achieve transformative outcomes.
Giving engineers structured customer touchpoints is the specific mechanism that unlocks AI's compounding innovation cycle.
2026-05-12
Thinking Machines' Interaction Models represent a scaled-up advance in voice AI, but the company overstates novelty to justify $2B of investment.
The real achievement is running fully-duplex at DeepSeek V4-Flash scale with video input — everything else is engineering catch-up dressed as innovation.
2026-05-18
Organizations that start with customer needs over technology capabilities extract significantly more value from AI investments.
Requiring every engineer to complete structured customer touchpoints is a concrete organizational mechanism — not a cultural platitude but a measurable practice.
4
Analyze
The agent produces a structured analysis
"The 2025 AI agent narrative is substantially overhyped — current agents are LLMs with basic function-calling bolted on, not truly autonomous systems, and enterprises are not yet ready for the governance, integration, and strategic clarity that real agentic deployment requires."
article_type: industry_shift reading_depth: skim familiarity: familiar response_angle: add_nuance
Key insight: The most consequential barrier to agentic AI adoption is not model capability — which IBM experts say is already sufficient — but enterprise readiness: the ability to expose internal APIs, organize proprietary data, and build governance frameworks before agents are deployed at scale.
5
Decide
The agent chooses its next action
Two tools are available: search_web (verify claims against external sources) or stop_here (the analysis is sufficient). The model reads the docstrings and picks.
search_web
Find external sources to corroborate claims. Used when the article is opinion-heavy or the thesis needs empirical grounding.
→ stop_here ←
The article's thesis is a well-established practitioner viewpoint. Searching would surface generic AI-hype discourse with no marginal value. The brief is complete.
Why: Although the article is classified as analysis with a critic stance rather than pure opinion, the reading_depth is skim and the core thesis is a well-established practitioner viewpoint. The key insight about enterprise readiness (APIs, data, governance) is the article's own framing — not an empirical claim requiring external corroboration.
6
Ground
Every claim is traced to a source
The agent cannot just assert. Every claim in the analysis must point to where it came from.
[1]
Current agents are LLMs with basic function-calling, not truly autonomous systems.
source: article
[2]
Enterprise readiness — exposing APIs, organizing proprietary data, building governance — is the primary barrier to agentic AI adoption.
source: article
[3]
99% of 1,000 enterprise developers surveyed by IBM and Morning Consult are exploring or developing AI agents.
source: article